Marius explains how TypeSafe AI's Jev differs from generative LLMs by outputting fast structured decisions rather than text. He demonstrates its three decision types—choice, score, and Noul—and shares informal custom comparison experiments evaluating speed, parallel evaluation, and classification consistency.
Original by mariusbuildsGetting startedIntermediate10 min 24 secPublished Source reviewed
Before you press play
What you’ll find in the video
Jev is a non-generative decision model designed to output structured outcomes—choice, score, or Noul—rather than conversational text.
Unlike autoregressive serial token generation, Jev evaluates multiple candidate options in parallel across a defined state.
In informal side-by-side experiments, Jev sustained consistent response latency across batches of questions where a conventional LLM timed out.
Worth knowing
Gemini-assisted video/transcript review. The author's sorting race and Spaceport Scanner demonstrations are informal experiments rather than controlled benchmarks, and decision accuracy varied across runs.